IJIRST –International Journal for Innovative Research in Science & Technology| Volume 3 | Issue 02 | July 2016 ISSN (online): 2349-6010
TLBO Optimization in Matching of Rotated Image Jyoti M. Tech Scholar Department of Computer Engineering Indus Institute of Engineering & Technology
Mr. Abhishek Bhatnagar Assistant Professor Department of Computer Engineering Indus Institute of Engineering & Technology
Abstract The main work of this research is the development of computational frameworks that deal with the problems arising in the construction of an image registration module. A novel algorithm for image features matching is used in terms of optimization algorithms. We have efficiently used the teacher learning based optimization algorithm (TLBO) to match image features by tuning the rotating angle of image. The complete work is divided into three modules: edge detection, features extraction and features matching. After studying previous work on this, the phase congruency method for the edge detection is used to avoid non uniform illumination problem occurred in detection of edges. The requirement of features extraction is that invariant matching points which remain same even after change in angle of rotation of image. So we used scale invariant features transform (SIFT) method to extract features and finally we purposed the optimized matching algorithm which can work for almost every type of image. Our features matching using TLBO algorithm tune the rotational angle of test image and set it to an optimum angle so that it can best match with the reference image. Keywords: TLBO algorithm, SIFT, GSA _______________________________________________________________________________________________________ I.
INTRODUCTION
Registration is a fundamental task in image processing used to match two or more pictures taken, for example, at different times, from different sensors, or from different viewpoints. Virtually all large systems which evaluate images require the registration of images, or a closely related operation, as an intermediate step. Specific examples of systems where image registration is a significant component include matching a target with a real time image of a scene for target recognition, monitoring global land usage using satellite images, matching stereo images to recover shape for autonomous navigation, and aligning images from different medical modalities for diagnosis. Image registration, sometimes called image alignment, is an important step for a great variety of applications such as remote sensing, medical imaging and multi-sensor fusion based target recognition. It is a prerequisite step prior to image fusion or image mosaic. Its purpose is to overlay two or more images of the same scene taken at different times, from different viewpoints and/or by different sensors. It is a fundamental image processing technique and is very useful in integrating information from different sensors, finding changes in images taken at different times, inferring threedimensional information from stereo images, and recognizing model-based objects. Registration can be performed either manually or automatically. In literature many features based image registration algorithm is widely used as due to contrast variation in image, pixel intensity varies and in these kind of cases features based extraction is efficient. Besides these the need of feature extraction in image registration requires the technique which is scale and affine transformation invariant. SIFT has been widely used in literature but the data length is quite big which requires more computational time. Iteratively matching algorithm have been used in literature which generally used genetic algorithm which suffers from falling into local maxima, so a more comprehensive method has to be proposed for iterative optimization. For these we have proposed a global optimization algorithm named gravitational search algorithm (GSA) along with SIFT features. A rotated image by 20 o angle is used as a test image and to avoid the edge variation due to intensity change, phase congruency algorithm is used followed by SIFT algorithm and GSA respectively. Next sections discuss the proposed algorithm and results. II. PROPOSED METHODOLOGY This work is related to image registration applied to almost every type of image. We have used phase congruency model discussed in previous chapter for edge detection. The selection of this model lies with the fact that images can vary in illumination of light and that poses threat to unidentified edges in image, resulting in skipping of that part in feature extraction. The complete work can be divided into three main modules as diagrammatically represented below in figure 1. Module 1: The edge detection in test image is done using PC algorithm which performs well over widely used canny edge detection doesn’t perform well in case of non-uniform illuminated image. The difference can be visible in the example pictures given below picked up from the Peter Koveski’s home page.
All rights reserved by www.ijirst.org
94
TLBO Optimization in Matching of Rotated Image (IJIRST/ Volume 3 / Issue 02/ 019)
Fig. 1: Modules of proposed work
Fig. 2 (a): input test image (b) PC edge detected map, Step and line features are marked equally well. Note how each whisker has been marked as a single feature. (c) canny edge detected map, note how edges are marked on each side of each whisker.
The mathematical formulation of 2-D image PC for our work is given as: first step is to convolve the normalized iris image I (x,y) with a bank of 2D log-Gabor filters with different orientations and scales. The Log-Gabor has a transfer function of the following form: −(log(ω⁄ω0 ))2 G(ω) = exp( ) 2(log(k⁄ω0 ))2 where ω0 is the filter’s center frequency. k⁄ω0 should be aconstant for vary ω0 . The 2D log-Gabor is constructed with the cross-section of the transfer function in the angular direction being a Gaussian function: −(θ − θ0 )2 G(θ) = exp( ) 2σθ 2 where θ0 represents the orientation of the filter, σθ is the standard deviation of this Gaussian function. Here we choose 6 even odd orientations and 4 scales. Let Mso and Mso denote the even-symmetric and odd symmetric filter at scale s and orientation o. The response vector can be got by: even odd [eso (x, y), oso (x, y)] = [I(x, y) ∗ Mso , I(x, y) ∗ Mso ] The amplitude of the response at a given scale and orientation can be computed by: Aso = √eso (x, y)2 + oso (x, y)2 And the phase angle is: ∅so (x, y) = atan(oso (x, y)/eso (x, y)) Let ∅so (x, y) , denotes the mean phase angle at orientation o, it can be estimated by:
All rights reserved by www.ijirst.org
95
TLBO Optimization in Matching of Rotated Image (IJIRST/ Volume 3 / Issue 02/ 019)
̅even ̅ odd (x, y), ∅ ∅ 0 0 (x, y) =
(∑s eso (x, y), ∑s oso (x, y),) √∑s eso (x, y)2 + ∑s oso (x, y)2
A sensitive phase deviation measure ∆∅so (x, y) , is used. ∆∅so (x, y) = cos (∅so (x, y) − ̅̅̅ ∅0 (x, y)) − |sin (∅so (x, y) − ̅̅̅ ∅0 (x, y))| Then the 2D phase congruency is calculated as follows: ∑o ∑s Aso (x, y)(∆∅so (x, y)) PC0 (x, y) = ∑o ∑s Aso (x, y)+∈ ∈is a very small positive real number, used to prevent division of zero, its value set to be 0.0001. According to Kovesi, using the magnitude of dot and cross products, the phase deviation can be calculated directly from the filter outputs, as ̅ even (x, y) + :Aso (x, y)∆∅so (x, y) = Aso (x, y) cos (∅so (x, y) − ̅̅̅ ∅0 (x, y)) − |sin (∅so (x, y) − ̅̅̅ ∅0 (x, y))| = (eso (x, y). ∅ 0 ̅odd ̅odd ̅even (x, y)) oso (x, y). ∅ 0 (x, y)) − (eso (x, y). ∅0 (x, y) + oso (x, y). ∅0 By this formula the phase congruency of image is calculated. The outcome of this is Module 2: Features extrication using shift invariant transform features is done in this module. The SIDT is explained in previous chapter so we won’t discuss again here. Module 3: Features mapping have equal importance as above steps. We have used a latest optimized matching algorithm i.e. teacher learner based optimisation algorithm (TLBO). The description of TLBO is done as: TLBO is based on knowledge transfer of teacher to student. Initially every student in a class is assigned a particular knowledge level. The fitness function to be evaluated is calculated for each different knowledge level and for which it comes out to be minimum, considered as teacher. Then teacher transfers his knowledge to students to raise their knowledge level by using the formula: Difference∗,D = r (Mnew,D − TF × M∗,D), Where TF is the teacher factor and it s values lies in between 1 and 2. After this objective function is called again and if any student has achieved the higher knowledge level than teacher then he will be awarded that position and process continues till iterations last. Steps are given below: Define the Optimization Problem and Initialize the Optimization Parameters: Initialize the population size (Pn), number of generations (Gn), number of design variables (Dn), and limits of design variables (UL, LL). Define the optimization problem as: minimize f (X) Subject to Xi ∈ xi = 1, 2, · · · , Dn where f (X) is the objective function, X is a vector for design variables such that L L ≤ xi ≤ UL. Initialize the population: Generate a random population according to the population size and number of design variables. For TLBO, the population size indicates the number of learners and the design variables indicate the subjects (i.e. courses) offered. This population is expressed as follows: x11 ⋯ x1D ⋱ ⋮ ] population = [ ⋮ xPn 1 ⋯ xPnD Teacher phase: Calculate the mean of the population column-wise, which will give the mean for the particular subject as M ∗,D = [m1, m2, · · · , mD]. The best solution will act as a teacher for that iteration X teacher = Xf(X)=min. The teacher will try to shift the mean from M∗,D towards X∗,teacher which will act as a new mean for the iteration. So, M new,D = Xteacher,D. The difference between two means is expressed as Difference∗,D = r (Mnew,D − TF × M∗,D), where, the value of TF is selected as 1 or 2. The obtained difference is added to the current solution to update its values using Xnew,D = Xold,D + Difference∗,D Accept Xnew if it gives better function value. Learner phase: As explained above, learners increase their knowledge with the help of their mutual interaction. The mathematical expression is explained as follows: Xnew,i = Xold,i + ri (xi − xj )iff(xi ) < f(xj ) Xnew,i = Xold,i + ri (xj − xi )iff(xi ) > f(xj ) Termination criterion: Stop if the maximum generation number is achieved; otherwise repeat from Step Teacher phase. In this chapter we have put a light on individual TLBO algorithm, how they behave and conserve using some mathematical equations. In next chapter we will discuss the method to cascade both of these algorithms so that results of optimisation can be improved. The reference image is rotated by some angle just for testing purpose. In TLBO, we optimize the optimum angle of rotation of the test image on the basis of minimizing mean square error in features of reference image and test image. The features
All rights reserved by www.ijirst.org
96
TLBO Optimization in Matching of Rotated Image (IJIRST/ Volume 3 / Issue 02/ 019)
descriptor obtained from the previous module for reference as well as test image is fed into matching module, so that a minimum difference between them can be reached. The fitness function to be minimized is the mean square error given as: sum(diff 2 ) MSE = size(features) The flow chart for the proposed work is given in figure. III. RESULTS The proposed work is implemented in MATLAB and we used medical DICOM images to test the image registration scheme. The input image is resized to 200* 200 pixels and is considered as the reference image which is rotated by 20 o to make a test image. Features are extracted from both images and converted into 1 dimensional vector. In matching we are using TLBO optimization. TLBO algorithm is used in our work and a total of 10 iterations for all 5 agents is done or in other words the objective function is called 10 times for an agent and the output of fitness function is plotted in figure 4. This is the plot of mean square error which is calculated in objective function. Since it is error, so it should be decreasing which is so in plot. For a good tuning of TLBO algorithm, the objective function plot must be decreasing and after little iteration it should settle to a minimum value with no further change. Around 40000 features are extracted out of 200*200 images. Since after rotation the black space is left behind and usually that doesnâ&#x20AC;&#x2122;t contain any information, so no feature points locations will be in that region. If any image is actually captured by rotating the lens of camera then this black space will not occur and a better features matching will also be visible. After using proposed matching algorithm the features points are matched more than simple matching algorithm whose result is shown in figure 4.
Features Matching without Optimisation
Fig. 4: features points matched in both images without optimization
Figure 5 shows the points matched in both images after rotating the image by optimized rotation angle and checking out points matching. After TLBO tuning the angle of -155.6571o is achieved or 24.3439o almost, which gives best matching results.
All rights reserved by www.ijirst.org
97
TLBO Optimization in Matching of Rotated Image (IJIRST/ Volume 3 / Issue 02/ 019)
Fig. 3: Flow chart of proposed work
An error of 4.3439o from the rotation angle at input is there. The normalized matching value calculated by number of points at same location for both cases is 0.2141 and 0.7326. Proposed algorithm is also tested on images other than DICOM images. A comparison bar graph for these is shown in figure 6. This bar graph shows that for the proposed work, normalized matching value is higher for every type of image.
All rights reserved by www.ijirst.org
98
TLBO Optimization in Matching of Rotated Image (IJIRST/ Volume 3 / Issue 02/ 019)
Features Matching with GSA Optimisation
Fig. 5: Points matched after GSA optimization
Without Optimization Pepper Image
With GSA Optimization
Beach image
Without Optimization
With GSA Optimization
All rights reserved by www.ijirst.org
99
TLBO Optimization in Matching of Rotated Image (IJIRST/ Volume 3 / Issue 02/ 019)
Fig. 6: normalized Matching for different types of images
IV. CONCLUSION We have used DICOM images for testing purpose but this is a generalized algorithm and can work for any type of input image whether that can be medical or satellite images. Unsupervised learning in the form of TLBO algorithm makes it generalize for every type of image. TLBO is an optimization algorithm which changes its agent’s position to find a place for minimum objective function value. In our case it will search for the optimum angle at which matching points between test and reference image are maximum. We considered mean square error between features points location extracted from the module 2 using SIFT. After optimization the rotational angle set to a minimum value for which more number of matching points exist and test image is aligned with the reference image by that optimum angle. The images used here are compressed to 200*200 so that execution time can be reduced. Comparison results with un-optimized solution shows that proposed work with TLBO optimization is better. REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22]
Medha V. Wyawahare, “Image Registration Techniques: An overview” International Journal of Signal Processing, Image Processing and Pattern Recognition, Vol. 2, No.3, September 2009. Shanmugapriya. S, “An intensity-based medical image Registration using genetic algorithm” Signal & Image Processing: An International Journal (SIPIJ) Vol.5, No.3, June 2014. Luiz Satoru Ochi, “Image Registration Using Genetic Algorithms” GECCO’08, July 12–16, 2008. Ch.Venkateswara Rao, “OPTIMIZATION OF AUTOMATIC IMAGE REGISTRATION ALGORITHMS AND CHARACTERIZATION”IEEE Trans. On Pattern Anal. And Machine Intell.2008. Andrea Valsecchi, “An Image Registration Approach using Genetic Algorithms” IEEE World Congress on Computational Intelligence June, 10-15, 2012. Siddharth Saxena, “ A Survey of Recent and Classical Image Registration Methods” International Journal of Signal Processing, Image Processing and Pattern Recognition Vol.7, No.4 (2014), pp.167-176. Feng Jin, “Image Registration Algorithm Using Mexican Hat Function-Based Operator and Grouped Feature Matching Strategy” e95576. doi:10.1371/journal.pone.0095576, 2014. Renbo Xia, “A robust feature-based registration method of multimodal image using phase congruency and coherent point drift” Pattern Recognition and Computer Vision, 2014. Peter Kovesi, “Phase Congruency Detects Corners and Edges” Multiple View Geometry in Computer Vision. Cam- bridge University Press (2000). P.Subha, “Automatic Feature Based Image Registration using DWT and SIFT” IJESC, 2014. Supriya S. Kothalkar, “Multimodal Image Registration using Contour let Transform” International Journal of Advanced Research in Computer and Communication Engineering,Vol. 3, Issue 3, March 2014. Dilipsinh Bheda, “A Study on Features Extraction Techniques for Image Mosaicing” International Journal of Innovative Research in Computer and Communication Engineering, Vol. 2, Issue 3, March 2014. Megha M Pandya , “Accurate Image Registration using SURF Algorithm by Increasing the Matching Points of Images” International Journal of Computer Science and Communication Engineering Volume 2 Issue 1 (February 2013 Issue). Rakesh Singhai , “Registration of Satellite Imagery Using Genetic Algorithm” Proceedings of the World Congress on Engineering 2012 Vol II WCE 2012, July 4 - 6, 2012. Mahmood Amintoosi, “Image Registration for Super-Resolution using SIFT Key-points”ICEE, 2009. Zaafouri Ahmed, “Satellite Images features Extraction using Phase Congruency model” IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.2, February 2009. Peter J. Myerscough, “ESTIMATING THE PHASE CONGRUENCY OF LOCALISED FREQUENCIES” Journal of Computer Vision Research, 2010. Xiaoyan Yuan, “Iris Feature Extraction Using 2D Phase Congruency” Proceedings of the Third International Conference on Information Technology and Applications, 2005. J. Krishna Chaithanya, “Feature Extraction of Satellite Images Using 2D Phase Congruency Measure” International Journal of Emerging Technology and Advanced Engineering, Volume 3, Issue 1, January 2013. J. Krishna Chaithanya, Dr. T. Rama Shri,” Feature Extraction of Satellite Images Using 2D Phase Congruency Measure”, International Journal of Emerging Technology and Advanced Engineering, Volume 3, Issue 1, January 2013 Andriy Myronenko and Xubo Song,” Point Set Registration: Coherent Point Drift”, IEEE Transactions On Pattern Analysis And Machine Intelligence, VOL. 32, NO. 12, DECEMBER 2010 Andrea Valsecchi and Sergio Damas,” An Image Registration Approach using Genetic Algorithms”, IEEE World Congress on Computational Intelligence June, 10-15, 2012 - Brisbane, Australia
All rights reserved by www.ijirst.org
100
TLBO Optimization in Matching of Rotated Image (IJIRST/ Volume 3 / Issue 02/ 019) [23] N.NagaRaju, T.Satya savitri, Ch.A.Swamy,” Image Registration Using Scale Invariant Feature Transform” International Journal of Scientific Engineering and Technology, Volume No.2, Issue No.7, 2013. [24] Haidawati Nasir, Vladimir Stankovic,” Image Registration For Super Resolution Using Scale Invariant Feature Transform, Belief Propagation And Random Sampling Consensus”, 18th European Signal Processing Conference (EUSIPCO-2010) [25] Gao Ting, Xu Yu, and Xu Ting-xin,” Multi-Scale Image Registration Algorithm based on Improved SIFT”, Journal of Multimedia, Vol. 8, No. 6, December 2013 [26] Samuel Cheng, Vladimir Stankovi´c and Lina Stankovi´c,” IMPROVED SIFT-BASED IMAGE REGISTRATION USING BELIEF PROPAGATION”, ICASSP 2009 [27] M.C. Morrone, J.R. Ross, ”Mach bands are phase dependent”, Nature, November 1986 [28] M.C. Morrone, R.A. owens, ’Feature detection from local energy’, Pattern Recognition Letters, 1987 [29] M.C. Morrone, D.C. Burr,’feture Detection in human vision: A phase dependent energy model’, proc. R. Soc. Lond. B. 1988
All rights reserved by www.ijirst.org
101